Novel Distributed Wavelet Transforms and
Routing Algorithms for Efficient Data Gathering
in Sensor Webs
PI: Antonio Ortega, USC
G. Shen S. Lee S.W. Lee S. Pattem A. Tu B. Krishnamachari
Viterbi School of Engineering, University of Southern California
M. Cheng S. Dolinar A. Kiely M. Klimesh
H. Xie
Jet Propulsion Laboratory, NASA
ESTC 2008  06/24/08 2
Efficient Sensor Web Communication Strategies Based on
Jointly Optimized Distributed Wavelet Transform and
Routing
Objectives
•Design algorithms that minimize energy consumption by
compressing correlated measurements as data is routed to the sink
•Enable nodes to reconfigure the network automatically by taking
into account variations in node characteristics
Example of a 2D field measured by a sensor web: (a) true
field, and reconstructed field using (b) distributed
wavelets or (c) quantized data with the same energy
consumption as in (b).
ESTC 2008  06/24/08 3
Efficient Sensor Web Communication Strategies Based on
Jointly Optimized Distributed Wavelet Transform and
Routing
Approach
•Develop data compression algorithms that exploit data correlation
Entropy coding, filter optimization, path merging, joint compression and routing, temporal coding,
compressed sensing
•Implement advances in networking and routing
Node selection, network initialization, routing optimization, link quality robustness, inclusion of
broadcast nodes, and automatic reconfigurability
•Test these new capabilities
In the lab, and in a sensor web of about 100 nodes in an outdoor realistic environment for an
extended period of time
Work to Date
•Data compression algorithms
Entropy coding, path merging, joint compression and routing, temporal coding, compressed sensing
•Advances in networking and routing
Node selection, routing optimization, inclusion of broadcasts
•In lab experiments
Preliminary experimental results for small, in lab networks
ESTC 2008  06/24/08 4
Presentation Outline
1.Related preAIST work
2.Data Compression
•Entropy Coding
•Spatiotemporal transforms and coding
•Spatiotemporal subsampling
•Compressed sensing
•Tree based 2D wavelet transforms
3.Networking
•Joint routing and transform optimization
•Inclusion of broadcast nodes
•Erasure Correcting Codes
4.Mote Implementation
•In lab implementation (tree based 2D wavelet)
ESTC 2008  06/24/08 5
PreAIST Joint Routing and Compression Technique
(Ciancio, Ortega, Pattem, Krishnamachari – USC)
•Unidirectional 5/3 lifting transform along routing paths [1]
•Heuristic for dealing with merged paths for 2D networks [2]
•Transform optimization per path
•Pros:
Unidirectional computation (no backward transmissions)
Pathwise transform optimization
Practical alternative to existing methods
•Cons:
Overhead from heuristic merging technique (not critically sampled)
Only exploits pathwise correlation
Optimization only pathwise, does not extend to 2D transforms
[1]. A Ciancio and A. Ortega, “A distributed wavelet compression algorithm for wireless multihop
sensor networks using lifting”, ICASSP’04.
[2]. A. Ciancio, S. Pattem, A. Ortega, B. Krishnamachari, “Energyefficient data representation
and routing for wireless sensor networks based on a distributed wavelet compression
algorithm”, IPSN’06.
ESTC 2008  06/24/08 6
Goal: Use entropy coding (data compression) to minimize cost of
transmitting values needed to compute a Discrete Wavelet Transform
(DWT) in the sensor web. This reduces the energy required to achieve
a given level of quality in the reconstructed data.
Motivation: Combined distributed DWT and entropy coding enables joint
compression of the data generated by different nodes as the
information accumulates over the routing path.
Results:
•Devised a general purpose entropy coding method for our transforms
•Details found in our NSTC 07 Paper [1][1]. G. Shen, et al, “A distributed wavelet approach for efficient information
representation and data gathering in sensor webs ”, NSTC’07.
Entropy Coding
(A. Kiely, M. Klimesh, H. Xie  JPL)
ESTC 2008  06/24/08 7
Goal:
•Combine temporal and spatial coding to minimize overall data transmission for
further energy reduction
•Consider both temporal and spatial correlation for node selectionMotivation:
•Most existing work focuses on spatial correlations only
•Data collected at each node exhibits high temporal correlation
•Temporal processing is local and far cheaper than spatial compression, and so
should be fully exploited to minimize the transmission cost
Two notable exceptions:
•Lightweight Temporal Coding [1]
•Distributed Predictive Coding [2][1]. T. Schoellhammer, B. Greenstein, E. Osterweil, M. Wimbrow, and D. Estrin, “Lightweight
temporal compression of microclimate datasets”, LCN'04.
[2]. A. Saxena and K. Rose, “Distributed predictive coding for spatiotemporally correlated
sources”, ISIT 2007.
SpatioTemporal Coding (H. Xie – JPL)
ESTC 2008  06/24/08 8
Key Observation:
•In data aggregationbased compression, data is transmitted through multiple hops along a
predefined routing path, and compressed jointly it flows towards the sink
•To encode data at a node for a given time instance, we can use all the information from the current
node / time instance along with data from previous nodes / time instances (see figure below)
Information flow along a 1D path in an aggregationbased data transmission system.
SpatioTemporal Coding
ESTC 2008  06/24/08 9
Assumptions
•Spatial routing path is established
•Latency introduced by local temporal
processing is tolerable
Approaches
•Separable wavelet transform
•Perform a single stage 3/5 reversible
integer DWT on data sequence at each
node to exploit temporal redundancy
•Perform spatial compression using
distributed wavelet transform and entropy
coding
•Adaptive filtering (work in progress)
•Each sample value is predicted from the
historical data
•The difference between the estimate and
the actual value is encoded and
transmitted
•The estimation error is also used to update
the filter weights
RateDistortion Performance
MSE Distortion
Average Transmission Cost
(Bits/Coefficient)
(a)
(b)
(c)
This ratedistortion graph shows the benefit of (a) Spatiotemporal
coding using 2D wavelet transform and (b) Spatial compression only,
compared to the baseline approach of (c) entropy coding quantized
sample differences (spatial only).
Example: 10bit source data as quantized version of 2D
secondorder Auto Regressive process with poles at
0.99e±j/64
SpatioTemporal Coding
ESTC 2008  06/24/08 10
•Spatiotemporal sampling patterns may lead to lower transmission cost
for same quality
•Benefits depend on spectral characteristics of data
SpatioTemporal Sampling
(S. Lee and A. Ortega – USC)
ESTC 2008  06/24/08 11
SpatioTemporal Sampling
Results
–
Real World Data : VTB data
[1]
There is max 2.6dB gain vs. temporalonly case in costPSNR sense.
0
0.005
0.01
0.015
0.02
0.025
0
0.05
0.1
2x temporal only
4x temporal only
6x temporal only
8x temporal only
2x spatiotemporal
4x spatiotemporal
6x spatiotemporal
8x spatiotemporal
0
0.005
0.01
0.015
0.02
0.025
20
30
40
50
60
70
80
90
100
2x temporal only
4x temporal only
6x temporal only
8x temporal only
2x spatiotemporal
4x spatiotemporal
6x spatiotemporal
8x spatiotemporal
[1] Jeongyeup Paek, Omprakash Gnawali KiYoung Jang, Daniel Nishimura, Ramesh Govindan, John Caffrey, Mazen Wahbeh, Sami Masri, A
Programmable Wireless Sensing System for Structural Monitoring, In: 4th World Conference on Structural Control and
Monitoring(4WCSCM), San Diego, CA, July 2006
ESTC 2008  06/24/08 12
Transforms and Routing
•Problem definition:
What data should be gathered from each node, how should it be
aggregated and transferred to the fusion node, how should it be
reconstructed
•Two major approaches:
Traditional techniques use data from all nodes, and reconstruct snapshots
of the state of field
a.Can handle any type of data
b.Exact reconstruction up to quantization error
We are exploring new techniques (i.e., compressed sensing, spatio
temporal sampling), where:
a.Some form of undersampling is used
b.Exact reconstruction only for classes of signals (sparse, bandlimited)
ESTC 2008  06/24/08 13
Compressed Sensing for Sensor Networks
(S. Lee, S. Pattem, B. Krishnamachari, A. Ortega)
Overview:
•Compressed Sensing (CS) is a technique capable of representing an N
length signal (which is Ksparse) using only M << N measurements
Goal:
•Design measurement matrices that lead to efficient routing while also
maintaining a high level of reconstruction quality
Experimental Observations:
•Our previous method [1] designed routing/measurement matrices that are
highly “incoherent” with the assumed signal basis (Fourier, Haar, etc)
However, correlation between coherence and reconstruction quality is low
•Spatial downsampling (DS) with CS is more efficient both in terms of
reconstruction quality and energy cost
[1]. G. Shen, et al, “A distributed wavelet approach for efficient information representation and data
gathering in sensor webs ”, NSTC’07.
ESTC 2008  06/24/08 14
Compressed Sensing for Sensor Networks
•DS consumes less energy for the same level of reconstruction quality
than DRP and SRP
AR data and DCT / Multilevel Haar basis
DS projection is highly incoherent with DCT basis and Haar basis.
Energy ratio vs. SNR of DS and SRP projections for AR data.
DRP is out of range due to very high energy cost.
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
5
10
15
20
25
30
35
40
45
ESTC 2008  06/24/08 15
Compressed Sensing for Sensor Networks
•CS with DS projection can provide a higher SNR at the same cost for
the low SNR region than 2D wavelet transform.
AR data and DCT basis for CS
CS has limited achievable SNR (unless the number of projections increases
significantly.)
CS with DCT basis vs. 2D wavelet transform.
ESTC 2008  06/24/08 16
Compressed Sensing for Sensor Networks
•Conclusions
Coherence is not accurate enough indicator for the reconstruction to
consider as design metric for measurement matrix
Downsampling with CS is efficient in terms of reconstruction quality and
energy consumption
•Future work
Extending to general topologies (currently using square grid)
ESTC 2008  06/24/08 17
Tree Based 2D Wavelet Transforms
(G. Shen, A. Ortega  USC)
Goal: Design and optimize a 2D transform
•Invertible, critically sampled, and unidirectional
•Exploits 2D correlation (not just pathwise, 1D correlation)
•Lifting filters arbitrary to permit filter optimizationMotivation:
•More decorrelation w/ 2D transform than pathwise transform [1]
•Existing 2D transforms require backwards transmissions [2]Our Proposed Method:
•Lifting transform along an arbitrary tree [3]
•Transform optimization via dynamic programming along a tree[1]. A Ciancio and A. Ortega, “A distributed wavelet compression algorithm for wireless multihop sensor
networks using lifting”, ICASSP’04.
[2]. R. Wagner, H. Choi, R. Baraniuk, V. Delouille, “Distributed wavelet transform for irregular sensor network
grids”, IEEE SSP’05.
[3]. G. Shen and A. Ortega, “Optimized distributed 2D transforms for irregularly sampled sensor network grids
using wavelet lifting”, ICASSP’08.
ESTC 2008  06/24/08 18
Tree Based 2D Wavelet Transforms
Lifting Transform Design (Split Design)
•Split sensors into even and odd nodes according to depth in tree
•Sequence of splitting trees across multilevels (derive T
j from Tj1
)
ESTC 2008  06/24/08 19
Tree Based 2D Wavelet Transforms
Lifting Transform Design (Filter Design)
•Can be computed in a variety of ways (planar regression, etc)
•We use simple averaging and smoothing ideas
Lifting Transform Computation
•Explicitly separate terms for parents
and children
•Permits unidirectional transform computation (no backwards tx)
Children of m in
splitting tree Tj
Parent of m in
splitting tree T
j
ESTC 2008  06/24/08 20
Optimized Unidirectional 2D Transforms on
Arbitrary Trees
Unidirectional Transform Computation
ESTC 2008  06/24/08 21
Tree Based 2D Wavelet Transforms
Transform Optimization (Cost Minimization for Fixed Distortion)
•Formulate as a Forward Dynamic Program
•Define the following quantities:
S = {1, 2, …, J} the set of coding schemes (levels of decomposition)
 the cost to transition from level i at n to level j at parent of n
 the optimal cost to arrive at level j at node n from its children
•We then have:
•The optimal solution is found using the results of Algorithm 1
ESTC 2008  06/24/08 22
Tree Based 2D Wavelet Transforms
Results (Highly Correlated AR2 Data, 100 Nodes)
ESTC 2008  06/24/08 23
Joint 2D Transform and Routing Optimization
(G. Shen, A. Ortega  USC)
Goal:
•Find routing trees that jointly optimize routing and compressionKey Observations:
•Shortest path routing trees (SPT) provide minimum total distance, not
always minimum per hop distance (for higher correlation)
•Hurts coding efficiency if data along tree is not well correlated
•Per hop correlation is higher along minimum spanning tree (MST)
•Tradeoff between low total multihop dist. and high per hop corr.
ESTC 2008  06/24/08 24
Joint 2D Transform and Routing Optimization
Optimization Problem:
•Find a spanning tree that minimizes total cost for a given distortionOur Proposed Method:
•Find an optimized combination of trees, i.e., MST and SPT [1]
•Useful approximation to true optimal tree
Number of spanning trees is very large (MatrixTree Theorem)
Can exploit tradeoff by combining min. cost routing (SPT) with high data
correlation tree (correlation based MST)
Two Main Methods:
•Search over all combinations of SPT and MST
•Start from an SPT, search from greatest depth nodes down to the sink
Exchange parent to that of parent in MST
If new tree is valid and cost is lower (for same distortion), change the parent from the
SPT to that in the MST
[1]. G. Shen and A. Ortega, “Joint routing and 2D transform optimization for irregular sensor
network grids using wavelet lifting”, IPSN’08.
ESTC 2008  06/24/08 25
Joint 2D Transform and Routing Optimization
Results for Uniform Network
Edges in MST are
more efficient
Heu. And Opt. differ
by only one edge
ESTC 2008  06/24/08 26
Joint 2D Transform and Routing Optimization
Results for Uniform Network (Continued)
Less than 1 dB diff.
between optimal and
heuristic trees
ESTC 2008  06/24/08 27
Broadcast Nodes
Goal:
•Design a communication
strategy to exploit these
extra communications
Motivation:
•Broadcast come for free
•The cost of the acquiring
messages at additional
nodes is negligible in many
realistic scenarios
Treestructure
communication paths
to the sink node
Same network with
possible additional “free”
communications arising
from the broadcast
nature of the links
Broadcast Nodes  Overview
Node communications are not directional; multiple nodes may be able to receive a
single transmission
ESTC 2008  06/24/08 28
Exploiting Broadcast Capability
(G. Shen, S. Pattem, A. Ortega – USC)
Goal:
•Utilize broadcast to enhance performance of our current systemMotivation:
•Broadcast allows nodes to transmit data to multiple neighbors
(more than their neighbors defined in a routing tree)
•Can use this to further decorrelate dataOur Proposed Method:
•Start with an initial routing tree T
•Exploit broadcast whenever possible along the tree
Augment the initial tree T to include broadcasts
Can even use more general graphs, forward data along T
ESTC 2008  06/24/08 29
Exploiting Broadcast Capability
Preliminary Broadcast Technique
•While forwarding along T, even node broadcasts at depth d can reach multiple
odd nodes at depth d+1, d+3, … , d1, d3, …
•Exploit these broadcasts by augmenting T into the graph T
A (see figure)
•Compute predicts along TA
and updates along T
Added links
ESTC 2008  06/24/08 30
Exploiting Broadcast Capability
Comparisons of Predict Computations
•Having more neighbors can produce lower predict energy
•Lower energy predicts require fewer bits to encode
Predict equations along TPredict equations along T
A
ESTC 2008  06/24/08 31
Exploiting Broadcast Capability
Preliminary Results
•Uniform 200 node network with highly correlated AR2 data
Green – broadcast links
Blue – tree links
ESTC 2008  06/24/08 32
Exploiting Broadcast Capability
Preliminary Results (Continued)
ESTC 2008  06/24/08 33
Erasure Correcting Codes  Link Quality
Considerations (M. Cheng, S. Dolinar  JPL)
Goal: Study and implement various erasurecorrecting codes for sensor
networks that achieve more reliable sensortosensor communications
Reduces number of retransmissions, resulting in energy savings
Allows for more robust system performance under varying degrees of link
quality
Motivation: Data transmissions between nodes are subject to erasures
with probability p, resulting in costly retransmissions that increase
overall energy consumption
Current Results:
•Developed software to generate and to simulate performance of:
Stateoftheart LT rateless codes
Blake’s new windowed erasurecorrecting codes
•Selected LT codes for implementation based on lower complexity and
greater maturity (but not yet implemented)
ESTC 2008  06/24/08 34
Mote Implementation
(S. Pattem, B. Krishnamachari – USC)
Implementation Related Goals:
•Provide inlab testbed for algorithm performance evaluation
•Identify and establish a real environment for real time algorithm
implementation
Current Results:
•Hardware implementation of invertible 2D wavelet
Tmote Sky devices (CC2420 radio)
•Storage and packetization
•Distributed and flexible operation (any given tree)
ESTC 2008  06/24/08 35
Mote Implementation
0
20
40
60
80
100
120
140
160
180
200
220
20
20.5
21
21.5
22
22.5
23
23.5
24
24.5
25
sample number
temperature
raw data
reconstruction
1
2
3
4
5
6
7
8
10
9
13
14
15
11
12
Base Station
Heat
Source
Heat
Source
0
20
40
60
80
100
120
140
160
180
200
220
20.5
21
21.5
22
22.5
23
23.5
24
24.5
25
25.5
sample number
temperature
raw data
reconstruction
node 12, 2 bits
node 12, 4 bits
Experimental setup
Result summary
Measurements are 16 bits
ESTC 2008  06/24/08 36
Mote Implementation
Results Show:
•Correctness of 2D wavelet implementation
•Tradeoff between cost and reconstruction quality
Looking Ahead
•For wide usage, need an architectural view that defines:
Where compression fits into standard network stack
Interactions between compression and networking components
Modules in compression component
Robustness mechanisms for distributed compression
ESTC 2008  06/24/08 37
Mote Implementation
Our Proposed Architecture
data,
robustness info
data, allow
processing
processed
data
order, role based
processing
structure
metric/goals
consistency
FORWARDING
ROUTING
LINK/NEIGHBOR
TABLE
COMPRESSION
CONFIGURATION
CLUSTERING
COMPUTATION
STORAGE
BIT REDUCTION &
PACKETIZATION
APPLICATION (SENSING)
MAC
measurements
order, role based
requirements
consistency
ESTC 2008  06/24/08 38
Future Work
•Data Compression Methods
Extensions of tree based 2D wavelets (i.e., “Treelets”)
Filter optimization methods
Extend temporal coding methods to tree based 2D wavelets
•Networking and Routing Methods
Node selection
Network initialization
Link quality robustness
Automatic reconfigurability
•Real Mote Implementation
Integrate entropy coding and erasure correcting codes
Develop software that handles an arbitrary sensor web topology
Test our algorithms in an outdoor realistic environment
ESTC 2008  06/24/08 39
Questions
???
Enter the password to open this PDF file:
File name:

File size:

Title:

Author:

Subject:

Keywords:

Creation Date:

Modification Date:

Creator:

PDF Producer:

PDF Version:

Page Count:

Preparing document for printing…
0%
Comments 0
Log in to post a comment